The basic reason, as mentioned in the text, is to be sure to form a relatively complete mapping or cover of those regions of X x A that actually occur. However, there could be good reasons for departing from this goal. For instance, X x A could have large regions that, though they do occur, they occur quite rarely compared with other regions. If classifier resources are limited, i.e., N is not very large, it might not be desirable to try to equalize resources over all regions--"holes" (lack of matching classifiers) in the frequent regions could result.

Such considerations aside, the first deletion technique results in all match sets having approximately the same number of classifiers in "microclassifier" terms. That is, the sum of the numerosities of macroclassifiers in any match set is approximately the same, regardless of the degree of generalization in those classifiers. The number of macroclassifiers could vary, however.


Why are you interested in allocating classifier resources approximately equally to the different match sets? I would guess that, mostly due to XCS generalization capabilities, different niches could require a different number of classifiers to be completely covered. What is your opinion on this point?